LLMCARE: early detection of cognitive impairment via transformer models enhanced by LLM-generated synthetic data

Alzheimer's disease and related dementias (ADRD) affect nearly five million older adults in the United States, yet more than half remain undiagnosed. Speech-based natural language processing (NLP) provides a scalable approach to identify early cognitive decline by detecting subtle linguistic markers that may precede clinical diagnosis.

This study aims to develop and evaluate a speech-based screening pipeline that integrates transformer-based embeddings with handcrafted linguistic features, incorporates synthetic augmentation using large language models (LLMs), and benchmarks unimodal and multimodal LLM classifiers. External validation was performed to assess generalizability to an MCI-only cohort.

Transcripts were obtained from the ADReSSo 2021 benchmark dataset (n = 237; derived from the Pitt Corpus, DementiaBank) and the DementiaBank Delaware corpus (n = 205; clinically diagnosed mild cognitive impairment [MCI] vs. controls). Audio was automatically transcribed using Amazon Web Services Transcribe (general model). Ten transformer models were evaluated under three fine-tuning strategies. A late-fusion model combined embeddings from the best-performing transformer with 110 linguistically derived features. Five LLMs (LLaMA-8B/70B, MedAlpaca-7B, Ministral-8B, GPT-4o) were fine-tuned to generate label-conditioned synthetic speech for data augmentation. Three multimodal LLMs (GPT-4o, Qwen-Omni, Phi-4) were tested in zero-shot and fine-tuned settings.

On the ADReSSo dataset, the fusion model achieved an F1-score of 83.32 (AUC = 89.48), outperforming both transformer-only and linguistic-only baselines. Augmentation with MedAlpaca-7B synthetic speech improved performance to F1 = 85.65 at 2 × scale, whereas higher augmentation volumes reduced gains. Fine-tuning improved unimodal LLM classifiers (e.g., MedAlpaca-7B, F1 = 47.73 → 78.69), while multimodal models demonstrated lower performance (Phi-4 = 71.59; GPT-4o omni = 67.57). On the Delaware corpus, the pipeline generalized to an MCI-only cohort, with the fusion model plus 1 × MedAlpaca-7B augmentation achieving F1 = 72.82 (AUC = 69.57).

Integrating transformer embeddings with handcrafted linguistic features enhances ADRD detection from speech. Distributionally aligned LLM-generated narratives provide effective but bounded augmentation, while current multimodal models remain limited. Crucially, validation on the Delaware corpus demonstrates that the proposed pipeline generalizes to early-stage impairment, supporting its potential as a scalable approach for clinically relevant early screening. All codes for LLMCARE are publicly available at: GitHub.